# Bank Churn using Nueral Network
The points distribution for this case is as follows:
Read the dataset
Drop the columns which are unique for all users like IDs (5points)
Distinguish the features and target variable(5points)
Divide the data set into training and test sets (5points)
Normalize the train and test data (10points)
Initialize & build the model. Identify the points of improvement and implement the same. Note that you need to demonstrate at least two models(the original and the improved one) and highlight the differences to complete this point. You can also demonstrate more models. (20points)
Predict the results using 0.5 as a threshold. Note that you need to first predict the probability and then predict classes using the given threshold (10points)
Print the Accuracy score and confusion matrix (5points)
In [50]:
# Start Python Imports
import math, time, random, datetime
# Data Manipulation
import numpy as np
import pandas as pd
# Visualization
import matplotlib.pyplot as plt
import missingno
import seaborn as sns
# Preprocessing
from sklearn.preprocessing import OneHotEncoder, LabelEncoder, label_binarize, StandardScaler
from sklearn.preprocessing import MinMaxScaler
# Machine learning
from sklearn.model_selection import train_test_split, cross_val_score
from sklearn import model_selection, tree, preprocessing, metrics, linear_model
from sklearn.svm import SVC
from sklearn.linear_model import Perceptron
from sklearn.svm import LinearSVC
from sklearn.ensemble import GradientBoostingClassifier
from sklearn.neighbors import KNeighborsClassifier
from sklearn.naive_bayes import GaussianNB
from sklearn.linear_model import LinearRegression, LogisticRegression, SGDClassifier
from sklearn.tree import DecisionTreeClassifier
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import confusion_matrix
# Use GridSearchCV to find the best parameters.
from sklearn.model_selection import GridSearchCV
from sklearn.metrics import confusion_matrix, accuracy_score, classification_report,roc_curve, auc
# Tensor Flow and Keras
import tensorflow as tf
import keras
from keras.models import Sequential
from keras.layers import Dense
from keras.layers import Dropout
In [51]:
# loading data set
data_raw = pd.read_csv("bank.csv")
In [52]:
data_raw.head(10)
Out[52]:
In [53]:
# making copy of raw data to do preprocesing
data = pd.DataFrame.copy(data_raw)
In [54]:
# shape of data set
data.shape
Out[54]:
In [55]:
# description of all variables
data.info()
In [56]:
# Check null data
data.isnull().sum()
Out[56]:
In [57]:
#Descripbe data
data.describe().T
Out[57]:
In [ ]:
In [58]:
data.head(10)
Out[58]:
In [59]:
# creating numerical, categorical & target columns lists
categorical_features = [features for features in data.columns if data[features].dtypes == "O"]
print(len(categorical_features),"no of categorical variables")
numerical_features = ["CreditScore", "Age", "Balance", "EstimatedSalary"]
print(len(numerical_features),"no of numerical variables")
discrete_features = ["Tenure","NumOfProducts","HasCrCard","IsActiveMember"]
print(len(discrete_features),"no of discrete variables")
target = ["Exited"]
print("target has",len(target),"variable")
In [60]:
# checking all unique values of categorical variable
for feature in categorical_features:
print(feature,"has",len(data[feature].value_counts()),"unique values","\n", data[feature].value_counts())
In [61]:
# checking all unique values of discrete variable
for feature in discrete_features:
print(feature,"has",len(data[feature].value_counts()),"unique values","\n", data[feature].value_counts())
In [62]:
# EDA
# need to check the distribution of output classes
# Output classes need to be balanced
data_EDA = pd.DataFrame(data)
sns.countplot(x="Exited", data=data_EDA)
plt.title("Distribution of output classes")
Out[62]:
In [63]:
# Univariate Analysis using histogram
# understanding more in histogram of caontinous variables
for features in numerical_features:
data_EDA = data.copy()
sns.distplot(data_EDA[features], hist=True, kde=True, rug=True)
plt.xlabel(features)
plt.xticks(rotation=90)
plt.title("Distribution of numerical variables")
plt.show()
In [64]:
#Age variable has right skewed distribution which means most customers are 25-50 years age group.
# Lot of customers have zero account balance.
In [65]:
# relationship between numerical variable vs target
for feature in numerical_features:
data_EDA=data.copy()
sns.barplot(x="Exited", y=feature, data=data_EDA, order=[1,0], palette=colors1)
plt.xlabel('Exited')
plt.xticks(rotation=90)
plt.ylabel(feature)
plt.title(feature)
plt.show()
In [66]:
#The average age group of exited customers is 45 years
# 650 is avearge credit score of existed customer
# 86000 is average balance & 100000 is average salary of exited cust
In [67]:
# Percentage of customer churn is ~20.4%
value_counts = pd.value_counts(data_EDA['Exited'])
plt.figure(figsize = (6,6))
value_counts.plot(kind = 'pie', explode = [0,0.1],autopct='%1.1f%%', shadow=True)
plt.title('Proportion of customer churned and retained')
plt.show()
value_counts
Out[67]:
In [74]:
# Multivariate Analysis
# checking relationship between categorical variables and target
for feature in categorical_features + discrete_features:
data_EDA=data.copy()
sns.countplot(x='Exited', hue=feature, data=data_EDA, order=[1,0])
plt.xlabel('Exited')
plt.xticks(rotation=90)
plt.ylabel('Count')
plt.title(feature)
plt.show()
In [75]:
#Germans and french people more likely to exit
# Women have more exit rate
# Customer who has credit card are more likely to leave
# Not active customer is leaving
# One year old customers are more likely to leave
In [76]:
# corelation plot
sns.pairplot(data_EDA, kind="scatter", hue="Exited", plot_kws=dict(s=120, edgecolor="white", linewidth=2.5))
plt.show()
In [71]:
plt.figure(figsize = (15,15))
sns.heatmap(data_EDA.corr(), annot = True, cmap = 'RdYlGn')
Out[71]:
In [77]:
# data preparation...creating independent and dependent features
# we do not consider Row Number and Customer Id as they are not needed .
X = data_EDA.iloc[:,3:-1].values
y = data_EDA.iloc[:,-1].values
In [78]:
X
Out[78]:
In [79]:
y
Out[79]:
In [80]:
# Label encoding Gender
from sklearn.preprocessing import LabelEncoder
le = LabelEncoder()
X[:, 2] = le.fit_transform(X[:, 2])
In [81]:
print(X)
In [82]:
# one hot encoding Geography
from sklearn.compose import ColumnTransformer
from sklearn.preprocessing import OneHotEncoder
ct = ColumnTransformer(transformers=[('encoder', OneHotEncoder(), [1])], remainder='passthrough')
X = np.array(ct.fit_transform(X))
In [83]:
X
Out[83]:
In [84]:
y
Out[84]:
In [86]:
# splitting train and test data
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X,y, test_size = 0.2, random_state = 0)
In [87]:
# Feature scaling using Standard scaler
from sklearn.preprocessing import StandardScaler
sc = StandardScaler()
X_train = sc.fit_transform(X_train)
X_test = sc.transform(X_test)
In [88]:
# Modeliing , building with ANN
Model 1 using Sequential Keras , RELU , batch size= 10 and EPoch 50¶
In [89]:
import tensorflow as tf
In [90]:
# initialize ANN with sequential model
ann = tf.keras.models.Sequential()
In [91]:
# adding input and first hidden layer use relu
ann.add(tf.keras.layers.Dense(units=6, activation='relu'))
In [92]:
# adding second hidden laer using relu
ann.add(tf.keras.layers.Dense(units = 6, activation = 'relu'))
In [93]:
# adding the output layer
ann.add(tf.keras.layers.Dense(units = 1, activation = 'sigmoid'))
In [94]:
# Compile the ANN
ann.compile(optimizer ='adam', loss = 'binary_crossentropy', metrics = ['accuracy'])
In [95]:
# Fitting ANN to train set with batch size 10 and epocs as 50
model_history = ann.fit(X_train, y_train,validation_split=0.33,batch_size = 10, epochs = 50)
In [96]:
# Visualize accuracy of model
plt.plot(model_history.history['accuracy'])
plt.plot(model_history.history['val_accuracy'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.show()
In [97]:
# vizualize the loss
plt.plot(model_history.history['loss'])
plt.plot(model_history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.show()
In [98]:
# Predict accuracy with threshold of 0.5
y_pred = ann.predict(X_test)
y_pred =(y_pred > 0.5)
In [99]:
# create confusion matrix
from sklearn.metrics import confusion_matrix, accuracy_score, classification_report
cm = confusion_matrix(y_test, y_pred)
print(cm)
print(accuracy_score(y_test, y_pred))
In [101]:
cr=metrics.classification_report(y_test,y_pred)
print(cr)
Model 2: Hyper-parameter tuning using Grid search with multiple learning rates and 5 fold validation¶
In [128]:
# tried with my laptop to run multiple batch sizez and epochs but it was taking too long to execute
# ended up with {'batch_size': 10, 'epochs': 150, 'optimizers': 'Adam'}
# 5 fold validation with n_jobs=-3 for faster speed.
from sklearn.model_selection import GridSearchCV
def build_classifier(lrs):
classifier = Sequential()
classifier.add(Dense(6, activation = "relu"))
classifier.add(Dense(1, activation = "sigmoid"))
classifier.compile(optimizer = keras.optimizers.Adam(lr = lrs), loss = "binary_crossentropy", metrics = ["accuracy"])
return classifier
classifier = KerasClassifier(build_fn = build_classifier, batch_size = 10, epochs = 150)
parameters = {'lrs':[ 0.001, 0.003]}
grid_search = GridSearchCV(estimator = classifier,
param_grid = parameters,
scoring = 'accuracy',
cv = 5 ,n_jobs=-3)
grid_search = grid_search.fit(X_train, y_train)
best_parameters = grid_search.best_params_
best_accuracy = grid_search.best_score_
In [129]:
display(best_parameters)
display(best_accuracy)
In [ ]:
# using above parameters to create new model
In [130]:
# initialize ANN with sequential model
ann_model2 = tf.keras.models.Sequential()
In [131]:
# adding input and first hidden layer use relu
ann_model2.add(tf.keras.layers.Dense(units=6, activation='relu'))
In [132]:
ann_model2.add(tf.keras.layers.Dense(units=1, activation='sigmoid'))
In [134]:
# Compile the ANN
sgd_optimizer = tf.keras.optimizers.SGD(lr=0.001)
ann_model2.compile(optimizer = sgd_optimizer, loss = 'binary_crossentropy', metrics = ['accuracy'])
In [138]:
# Fitting ANN to train set with batch size 10 and epocs as 150
model2_history = ann_model2.fit(X_train, y_train,validation_split=0.33,batch_size = 10, epochs = 150)
In [139]:
# Visualize accuracy of model
plt.plot(model2_history.history['accuracy'])
plt.plot(model2_history.history['val_accuracy'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper left')
plt.show()
In [140]:
# vizualize the loss
plt.plot(model2_history.history['loss'])
plt.plot(model2_history.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'validation'], loc='upper right')
plt.show()
In [141]:
# Predict accuracy with threshold of 0.5
y_pred = ann_model2.predict(X_test)
y_pred =(y_pred > 0.5)
In [142]:
# create confusion matrix
from sklearn.metrics import confusion_matrix, accuracy_score, classification_report
cm = confusion_matrix(y_test, y_pred)
print(cm)
print(accuracy_score(y_test, y_pred))
In [143]:
cr=metrics.classification_report(y_test,y_pred)
print(cr)
In [ ]:
# with the change in the hyper params we are able to improve the model accuracy from 0.852 to 0.862
#Germans and french people more likely to exit
# Women have more exit rate
# Customer who has credit card are more likely to leave
# Not active customer is leaving
# One year old customers are more likely to leave
#The average age group of exited customers is 45 years
# 650 is avearge credit score of existed customer
# 86000 is average balance & 100000 is average salary of exited cust.
# Lot of customers have zero account balance.